A fault prediction method for oil-immersed transformer and related device

Abstract

The application provides a fault prediction method and related device of an oil-immersed transformer, and belongs to the technical field of transformer fault prediction. According to the fault stage division result of the oil-immersed transformer, the fault stage in which the current oil-immersed transformer is located is determined; the first data is retained to evolve the law, the second data is extracted to reflect the sensitivity feature of the oil-immersed transformer fault, the third data is extracted to reflect the physical essence feature of the oil-immersed transformer fault, and the first feature vector, the second feature vector and the third feature vector are obtained correspondingly; according to the fault stage in which the current oil-immersed transformer is located, the first feature vector, the second feature vector and the third feature vector are fused to obtain a fused channel feature vector; and the fused channel feature vector is input into an improved DCCNN model with determined parameters to obtain a fault prediction result of the oil-immersed transformer. The application solves the problem that the accuracy of the fault prediction of the oil-immersed transformer is not high.

Description

Technical Field

[0001] This invention belongs to the field of transformer fault prediction technology, specifically relating to a fault prediction method and related device for oil-immersed transformers. Background Technology

[0002] Oil-immersed transformers are core equipment in power systems, and their safe and stable operation directly affects the reliability of power grid supply and socio-economic security. Under long-term electrical, thermal, and mechanical stress, the insulating oil and solid insulating materials inside the transformer decompose, producing characteristic gases such as hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2). Dissolved gas analysis (DGA) is an internationally recognized core technology for predicting internal transformer faults.

[0003] Currently, fault prediction methods for oil-immersed transformers are mainly divided into two categories: The first is the traditional DGA method (such as the three-ratio method and the Rogers ratio method), which relies on human experience and is not only greatly affected by subjective factors but also cannot adapt to complex fault modes. The second is machine learning / deep learning-based methods, such as Support Vector Machine (SVM) and Convolutional Neural Network (CNN). Although these methods improve the fault prediction accuracy of oil-immersed transformers, they have the following significant engineering adaptability defects in practical power engineering applications:

[0004] (1) Sample processing is detached from the actual power field: Sample balance is achieved only through oversampling, without considering the boundary stability and risk level differences of power field samples. High-risk noise samples (such as abnormal data close to the fault threshold) are prone to model bias, and there is a lack of sample prediction decision evaluation mechanism, which is not compatible with the stringent requirements of power grid operation and maintenance for data reliability.

[0005] (2) Feature extraction is disconnected from the fault evolution mechanism: Transformer faults have a three-stage evolution characteristic of "latency-development-fault" (in accordance with the "Guideline for Condition Evaluation of Power Transformers" DL / T 573-2021). The gas characteristics and physical properties of different stages are significantly different (e.g., H2 dominates in the latent stage and C2H2 surges in the fault stage). Existing methods use a fixed feature extraction structure, which cannot adapt to the stage changes, resulting in insufficient feature extraction targeting.

[0006] (3) The channel structure is too simple and does not integrate the core physical characteristics of the power system: It only relies on one-dimensional DGA data or two-dimensional image features and does not include core physical parameters such as transformer hot spot temperature, partial discharge quantity and insulating oil dielectric loss factor. These parameters can directly reflect the insulation status and thermal stability of the equipment and are the key basis for power engineering fault prediction. The absence of these parameters will cause the feature dimensions to deviate from the essence of equipment operation, which will affect the accuracy of transformer fault prediction.

[0007] (4) The output information cannot support operation and maintenance decisions: It only outputs the fault type and lacks key information such as prediction reliability and gas contribution. It cannot provide the operation and maintenance personnel of the power system with accurate maintenance direction, resulting in delayed operation and maintenance response, which does not meet the practical requirements of the "Guidelines for Condition-Based Maintenance of Power Equipment" DL / T 304-2021.

[0008] (5) Low efficiency of hyperparameter tuning: Traditional grid search and random search methods have high computational costs and long iteration cycles, making it difficult to meet the engineering requirements of power systems for real-time performance, stability and less downtime of prediction models. Summary of the Invention

[0009] The purpose of this invention is to provide a fault prediction method and related apparatus for oil-immersed transformers, which solves the problem of low accuracy in fault prediction of oil-immersed transformers in the prior art.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a fault prediction method for an oil-immersed transformer, comprising the following steps:

[0012] Real-time acquisition of dissolved gas analysis data in oil of oil-immersed transformers, as well as fault sensitivity ratio characteristic data and physical characteristic data, are recorded as the first data, the second data and the third data, respectively, which together form sample data, and the sample data includes normal samples and abnormal samples; and the fault stages of oil-immersed transformers are divided according to the sample data, and the current fault stage of oil-immersed transformers is determined according to the fault stage division results of oil-immersed transformers.

[0013] The first data retains the time evolution law, the second data extracts the sensitivity features reflecting the faults of oil-immersed transformers, and the third data extracts the physical essence features reflecting the faults of oil-immersed transformers, thus obtaining the first feature vector, the second feature vector, and the third feature vector.

[0014] Based on the current fault stage of the oil-immersed transformer, the first feature vector, the second feature vector, and the third feature vector are fused to obtain the fused channel feature vector.

[0015] The fused channel feature vectors are input into an improved DCCNN model with predetermined parameters to obtain fault prediction results for oil-immersed transformers. The improved DCCNN model is obtained by adding dilated convolution modules between the elements of the standard convolution kernel of the DCCNN model. The dilated convolution modules are used to expand the receptive field, and the parameters are determined based on the first and third data.

[0016] A further improvement of this invention is that, before obtaining the first feature vector, the second feature vector, and the third feature vector, the abnormal samples in the first, second, and third data are subjected to sample equalization processing to obtain the equalized samples. This process involves retaining the evolution law of the first data, extracting the sensitivity features reflecting the faults of the oil-immersed transformer from the second data, and extracting the physical essence features reflecting the faults of the oil-immersed transformer from the third data.

[0017] A further improvement of the present invention is that the abnormal samples in the first data, the second data and the third data are subjected to sample equalization processing. After obtaining the equalized samples, the stability of the equalized samples is judged. If the stability of the equalized samples does not meet the conditions, the equalized samples are subjected to sample equalization processing again.

[0018] A further improvement of the present invention is that the abnormal samples include faulty samples and high-risk samples;

[0019] The process of equalizing outlier samples in the first, second, and third data sets specifically involves:

[0020] The fault samples in the first, second, and third data are subjected to SMOTE oversampling for sample equalization to obtain synthetic samples.

[0021] The noise in the synthesized samples was processed using the IDBCS method for sample equalization.

[0022] High-risk samples in the first, second, and third datasets are weighted to achieve sample equalization.

[0023] A further improvement of this invention is that the calculation formula for fusing the first feature vector, the second feature vector, and the third feature vector is as follows:

[0024]

[0025] in, This is the fused channel feature vector. for function, The first eigenvector, This is the second feature vector. The third eigenvector, The weights of the first feature vector. The weights of the second feature vector. The weights are the third feature vector.

[0026] A further improvement of the present invention is that the parameters are determined based on the first data and the third data, specifically: an adaptive feature extraction is performed on the first data and the third data using a fault identification model to determine the parameters of the improved DCCNN model.

[0027] A further improvement of this invention is that the calculation formula for the dilated convolution module is:

[0028]

[0029] in, For the index of the convolutional layer, For the first convolutional layer feature maps at location The activation value on For the first The kernel weights of the convolutional layer For input data, For the first The bias term of the convolutional layer, Let σ be the void ratio, and σ be the nonlinear activation function. For the first Row index of the convolution kernel of the convolutional layer. For the first Column index of the convolution kernel of the convolutional layer. and These are the x and y coordinates on the feature map.

[0030] Secondly, the present invention provides a fault prediction system for an oil-immersed transformer, comprising:

[0031] The data acquisition module is used to acquire dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data in real time for oil-immersed transformers. These are denoted as the first data, the second data, and the third data, respectively, and together they form sample data. The sample data includes normal samples and abnormal samples. The module then divides the fault stages of the oil-immersed transformer based on the sample data and determines the current fault stage of the oil-immersed transformer based on the fault stage division results.

[0032] The feature vector determination module is used to retain the time evolution law of the first data, extract the sensitivity features reflecting the fault of the oil-immersed transformer from the second data, and extract the physical essence features reflecting the fault of the oil-immersed transformer from the third data, thereby obtaining the first feature vector, the second feature vector and the third feature vector.

[0033] The fusion module is used to fuse the first feature vector, the second feature vector, and the third feature vector according to the current fault stage of the oil-immersed transformer to obtain the fused channel feature vector.

[0034] The prediction module is used to input the fused channel feature vector into an improved DCCNN model with determined parameters to obtain the fault prediction result of the oil-immersed transformer; wherein, the improved DCCNN model is obtained by adding a dilated convolution module between the elements of the standard convolution kernel of the DCCNN model, the dilated convolution module is used to expand the receptive field, and the parameters are determined based on the first data and the third data.

[0035] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the fault prediction method for oil-immersed transformers described above.

[0036] Fourthly, the present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the fault prediction method for oil-immersed transformers described above.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The fault prediction method for oil-immersed transformers proposed in this invention considers not only dissolved gas analysis data in the transformer oil, but also fault sensitivity ratio characteristic data and physical characteristic data. Considering only one type of data can lead to low accuracy in subsequent transformer fault prediction; comprehensively considering all three types of data improves the accuracy of subsequent fault prediction. Furthermore, this invention divides the fault stages of the oil-immersed transformer based on real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data. This operation makes it easier for the subsequently improved DCCNN model to capture the true fault trend, thereby improving the accuracy of subsequent transformer fault prediction. Moreover, this invention fuses the first, second, and third feature vectors according to the current fault stage of the oil-immersed transformer to obtain a fused channel feature vector. This operation reduces the risk of overfitting in the improved DCCNN model, thereby improving the accuracy of fault prediction for oil-immersed transformers. Furthermore, this invention improves the DCCNN model by adding dilated convolutional modules between the elements of the standard convolutional kernel of the DCCNN model, thereby capturing long-range dependencies and multi-scale features, and thus improving the accuracy of fault prediction for oil-immersed transformers.

[0039] Furthermore, this invention discloses the following steps: before obtaining the first feature vector, the second feature vector, and the third feature vector, the abnormal samples in the first, second, and third data are subjected to sample equalization processing to obtain the equalized samples. This operation can reduce the bias of the later improved DCCNN model, reduce the false negative rate, and thus improve the accuracy of transformer fault prediction. Attached Figure Description

[0040] Figure 1 This is a flowchart of the fault prediction method for oil-immersed transformers according to the present invention;

[0041] Figure 2 This is a schematic diagram of the fault prediction system for the oil-immersed transformer of the present invention;

[0042] Figure 3 This is a flowchart of the fault prediction method for oil-immersed transformers in Embodiment 4 of the present invention;

[0043] Figure 4 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0044] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0045] Example 1:

[0046] The flowchart of the fault prediction method for oil-immersed transformers of the present invention is as follows: Figure 1 As shown, the fault prediction method for oil-immersed transformers of the present invention includes the following steps:

[0047] S1. Real-time acquisition of dissolved gas analysis data in oil of oil-immersed transformer, as well as fault sensitivity ratio characteristic data and physical characteristic data, which are respectively recorded as first data, second data and third data, together forming sample data, and the sample data includes normal samples and abnormal samples; and the fault stage of oil-immersed transformer is divided according to the sample data, and the current fault stage of oil-immersed transformer is determined according to the fault stage division results of oil-immersed transformer.

[0048] S2. The time evolution law of the first data is retained, the sensitivity features reflecting the fault of the oil-immersed transformer are extracted from the second data, and the physical essence features reflecting the fault of the oil-immersed transformer are extracted from the third data, thus obtaining the first feature vector, the second feature vector and the third feature vector.

[0049] S3. Based on the current fault stage of the oil-immersed transformer, the first feature vector, the second feature vector, and the third feature vector are fused to obtain the fused channel feature vector.

[0050] S4. Input the fused channel feature vector into the improved DCCNN model with determined parameters to obtain the fault prediction result of the oil-immersed transformer; wherein, the improved DCCNN model is obtained by adding dilated convolution modules between the elements of the standard convolution kernel of the DCCNN model, the dilated convolution modules are used to expand the receptive field, and the parameters are determined based on the first data and the third data.

[0051] Example 2:

[0052] A schematic diagram of the fault prediction system for the oil-immersed transformer of this invention is shown below. Figure 2 As shown, the fault prediction system for the oil-immersed transformer of the present invention includes:

[0053] The data acquisition module is used to acquire dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data in real time for oil-immersed transformers. These are denoted as the first data, the second data, and the third data, respectively, and together they form sample data. The sample data includes normal samples and abnormal samples. The module then divides the fault stages of the oil-immersed transformer based on the sample data and determines the current fault stage of the oil-immersed transformer based on the fault stage division results.

[0054] The feature vector determination module is used to retain the time evolution law of the first data, extract the sensitivity features reflecting the fault of the oil-immersed transformer from the second data, and extract the physical essence features reflecting the fault of the oil-immersed transformer from the third data, thereby obtaining the first feature vector, the second feature vector and the third feature vector.

[0055] The fusion module is used to fuse the first feature vector, the second feature vector, and the third feature vector according to the current fault stage of the oil-immersed transformer to obtain the fused channel feature vector.

[0056] The prediction module is used to input the fused channel feature vector into an improved DCCNN model with determined parameters to obtain the fault prediction result of the oil-immersed transformer; wherein, the improved DCCNN model is obtained by adding a dilated convolution module between the elements of the standard convolution kernel of the DCCNN model, the dilated convolution module is used to expand the receptive field, and the parameters are determined based on the first data and the third data.

[0057] Example 3:

[0058] The fault prediction method for oil-immersed transformers of the present invention includes the following steps:

[0059] S1. Real-time acquisition of dissolved gas analysis data in oil of oil-immersed transformer, as well as fault sensitivity ratio characteristic data and physical characteristic data, which are respectively recorded as first data, second data and third data, together forming sample data, and the sample data includes normal samples and abnormal samples; and the fault stage of oil-immersed transformer is divided according to the sample data, and the current fault stage of oil-immersed transformer is determined according to the fault stage division results of oil-immersed transformer.

[0060] The fault stage division results of the oil-immersed transformer in this step include the fault latency stage, the fault development stage, and the fault occurrence stage.

[0061] S2. The time evolution law of the first data is retained, the sensitivity features reflecting the fault of the oil-immersed transformer are extracted from the second data, and the physical essence features reflecting the fault of the oil-immersed transformer are extracted from the third data, thus obtaining the first feature vector, the second feature vector and the third feature vector.

[0062] The time evolution pattern of the first data is preserved, the sensitivity features reflecting oil-immersed transformer faults are extracted from the second data, and the physical essence features reflecting oil-immersed transformer faults are extracted from the third data. Before obtaining the first, second, and third feature vectors, the abnormal samples in the first, second, and third data are subjected to sample equalization processing to obtain the equalized samples. Abnormal samples include fault samples and high-risk samples.

[0063] The outlier samples in the first, second, and third data sets are subjected to sample equalization. After obtaining the equalized samples, the stability of the equalized samples is judged. If the stability of the equalized samples does not meet the conditions, the equalization process is repeated.

[0064] This step involves performing sample equalization on outlier samples in the first, second, and third datasets, specifically as follows:

[0065] The fault samples in the first, second, and third data are subjected to SMOTE oversampling for sample equalization to obtain synthetic samples.

[0066] The noise in the synthesized samples was processed using the IDBCS method for sample equalization.

[0067] High-risk samples in the first, second, and third datasets are weighted to achieve sample equalization.

[0068] S3. Based on the current fault stage of the oil-immersed transformer, the first feature vector, the second feature vector, and the third feature vector are fused to obtain the fused channel feature vector.

[0069] Based on the current fault stage of the oil-immersed transformer, the calculation formula for fusing the first, second, and third eigenvectors is as follows:

[0070]

[0071] in, This is the fused channel feature vector. for function, The first eigenvector, This is the second feature vector. The third eigenvector, The weights of the first feature vector. The weights of the second feature vector. The weights are the third feature vector.

[0072] S4. Input the fused channel feature vectors into the improved DCCNN model with determined parameters to obtain the fault prediction results of the oil-immersed transformer.

[0073] The improved DCCNN model in this step is obtained by adding dilated convolutional modules between the elements of the standard convolutional kernel of the DCCNN model. The dilated convolutional modules are used to expand the receptive field, and the parameters are determined based on the first and third data.

[0074] The formula for calculating the dilated convolution module is:

[0075]

[0076] in, For the index of the convolutional layer, For the first convolutional layer feature maps at location The activation value on For the first The kernel weights of the convolutional layer For input data, For the first The bias term of the convolutional layer, Let σ be the void ratio, and σ be the nonlinear activation function. For the first Row index of the convolution kernel of the convolutional layer. For the first Column index of the convolution kernel of the convolutional layer. and These are the x and y coordinates on the feature map.

[0077] In this step, the parameters are determined based on the first and third data. Specifically, the fault identification model is used to perform adaptive feature extraction on the first and third data to determine the parameters of the improved DCCNN model.

[0078] Example 4:

[0079] The method of the present invention will be described in detail below. The flowchart of the fault prediction method for oil-immersed transformers of the present invention is as follows: Figure 3 As shown, the fault prediction method for oil-immersed transformers of the present invention includes the following steps:

[0080] S1. Real-time acquisition of dissolved gas analysis data in oil-immersed transformer oil, as well as fault sensitivity ratio characteristic data and physical characteristic data, which are respectively recorded as the first data, the second data and the third data, together forming sample data, and the sample data includes normal samples and abnormal samples.

[0081] The following is a detailed explanation of step S1:

[0082] First, dissolved gas analysis (DGA) data (referred to as first data), fault sensitivity ratio characteristic data (referred to as second data), and physical characteristic data (referred to as third data) in oil-immersed transformer oil are acquired in real time as sample data. The real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data (referred to as ratio characteristic data), and physical characteristic data in oil-immersed transformer oil all include normal samples and abnormal samples.

[0083] The following provides a detailed explanation of the real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data in oil-immersed transformer oil:

[0084] A. Dissolved gas analysis data in oil-immersed transformer oil.

[0085] The real-time concentration values ​​(in μL / L) of H2, CH4, C2H4, C2H6 and C2H2 were collected by an online oil chromatography monitoring device at a sampling frequency of 1 time / hour and a data accuracy of ±0.1 μL / L, providing a core basis for fault type identification.

[0086] B. Fault sensitivity ratio characteristic data.

[0087] Based on the industry standard for power fault prediction, a 16-dimensional fault sensitivity ratio feature is derived (as shown in Table 1), covering the proportion of gas components and the total proportion, which directly corresponds to the core logic of the traditional three-ratio method and Rogers ratio method, and realizes the engineering connection with traditional prediction methods.

[0088] Table 1. 16-Dimensional Fault Sensitivity Ratio Characteristic Data

[0089]

[0090] In Table 1, TCH = CH4 + C2H2 + C2H4 + C2H6, and D = CH4 + C2H2 + C2H4.

[0091] C. Physical characteristic data.

[0092] By collecting transformer hot spot temperature through fiber optic temperature sensors, partial discharge quantity through ultrasonic sensors, and dielectric loss factor of insulating oil through dielectric loss testers, these parameters directly reflect the insulation status, thermal stability, and discharge intensity of the transformer, and are core indicators for the condition evaluation of power engineering.

[0093] S2. The time evolution law of the first data is retained, the sensitivity features reflecting the fault of the oil-immersed transformer are extracted from the second data, and the physical essence features reflecting the fault of the oil-immersed transformer are extracted from the third data, thus obtaining the first feature vector, the second feature vector and the third feature vector.

[0094] The specific implementation process of step S2 is as follows: the dissolved gas analysis data, fault sensitivity ratio characteristic data and physical characteristic data acquired in real time in the oil-immersed transformer oil are input into the GAF image channel, ratio characteristic channel and physical characteristic channel respectively to obtain the GAF image channel feature vector (called the first feature vector), the ratio characteristic channel feature vector (called the second feature vector) and the physical characteristic channel feature vector (called the third feature vector).

[0095] In this step, the GAF image channel is used to preserve the temporal evolution of dissolved gas analysis data in oil-immersed transformers. The ratio feature channel is used to reflect the sensitivity characteristics of faults in oil-immersed transformers. The physical feature channel is used to reflect the essential physical characteristics of faults in oil-immersed transformers.

[0096] Before obtaining the feature vectors for the GAF image channel, ratio feature channel, and physical feature channel, respectively, the real-time acquired dissolved gas analysis data, fault sensitivity ratio feature data, and physical feature data in the oil-immersed transformer oil are input into the GAF image channel, ratio feature channel, and physical feature channel, respectively. Then, sample equalization processing is performed on the abnormal samples in the real-time acquired dissolved gas analysis data, fault sensitivity ratio feature data, and physical feature data in the oil-immersed transformer oil to obtain the equalized samples. Abnormal samples include fault samples and high-risk samples.

[0097] The following is a detailed explanation of faulty samples and high-risk samples:

[0098] Fault samples refer to samples that reflect the actual fault state and are the "effective abnormal samples" that the model needs to learn.

[0099] High-risk samples refer to samples that, although they do not meet the failure criteria, have a high risk level and are likely to develop into failures.

[0100] This step involves sample equalization processing of abnormal samples in the real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data of oil-immersed transformer oil. Specifically:

[0101] a. The SMOTE oversampling method is used to perform sample equalization processing on the real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data and physical characteristic data of oil-immersed transformer oil (such as partial discharge, which accounts for only 12.5%) to obtain a synthetic sample.

[0102] b. The noise in the synthetic samples is equalized using the IDBCS method (in... Figure 3 (represented by IDBCS noise cleanup in Chinese).

[0103] c. High-risk samples in the real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data of oil-immersed transformers are subjected to sample equalization processing using a weighting reduction method. Figure 3 (High-risk samples are represented by reduced weights).

[0104] The formula for calculating the synthetic sample in a is:

[0105]

[0106] in, For the synthetic sample, rand (0,1) is a random number between [0,1]. for of K One of the nearest neighbor samples, For the minority class sample set, the first One sample.

[0107] The following is a detailed explanation of the IDBCS method for sample equalization:

[0108] IDBCS (Improved Probability Density-Based Data Cleaning Strategy) is a noise removal algorithm designed in this embodiment for the unbalanced sample scenario of transformer DGA. Its core principle is to clean the noise of the synthetic samples generated by SMOTE oversampling, rather than the original data, so as to avoid the loss of fault information and optimize the sample balance effect.

[0109] The principles and implementation steps of the IDBCS method are explained below:

[0110] a. Determine the majority class boundary samples M brdFirst, find the potential boundary samples of the majority class that contain minority class samples in the K nearest neighbors, calculate their probability density relative to the minority class samples, and filter out the boundary samples of the majority class, but do not delete the original samples.

[0111] For a majority class sample, if its K nearest neighbors include minority class samples, then the sample is classified as a potential boundary sample. M over For any xj M over Its density formula relative to the minority class samples is:

[0112]

[0113] in, Mmin For the collection of minority class samples, express Mmin The first in There are samples, where ω is the set of minority class samples. Mmin Statistical characteristics ω=ασd ,in, σ and d for Mmin The variance and mean of all samples in the dataset. α Hyperparameters α The typical value range for is [0.8, 2.0]. p rbf is the radial basis function of the Gaussian kernel, | Mmin | indicates the number of minority class samples.

[0114] The following explains minority class samples and boundary samples:

[0115] Minority samples are a type of faulty sample. Minority samples are those samples or a few samples that are far fewer in number than other categories (majority categories) among all sample categories.

[0116] Boundary samples refer to samples that are in the ambiguous region between "normal" and "abnormal" categories, and whose category classification is unclear (neither purely normal nor purely abnormal).

[0117] For any x Mover Determine the majority class samples to be deleted, i.e., the boundary samples. Mbrd The calculation formula is:

[0118]

[0119] Among them, | Mover | indicates Mover The number of samples in the middle.

[0120] b. Calculate the synthetic sample density: For the synthetic minority class sample set generated by SMOTE, calculate the density of each synthetic sample relative to the majority class boundary sample. Mbrd The probability density function is calculated using the following formula:

[0121]

[0122] Among them, | Mbrd | indicates Mbrd The number of samples in the middle.

[0123] c. Remove synthetic noise samples: Delete synthetic samples located in high-density regions of the majority class boundary to avoid noise samples interfering with model training and ensure the quality of synthetic samples. The calculation formula is as follows:

[0124]

[0125] Among them, | Msyn | indicates Msyn The number of samples in the middle.

[0126] The following is a detailed explanation of high-risk samples and the process of reducing the weight of high-risk samples:

[0127] The distance between outlier samples and the decision boundary is calculated using Euclidean distance. D ,when D <0.1 D max ( D max When the maximum distribution distance of abnormal samples is reached, samples are identified as high-risk samples (such as abnormal data close to the fault threshold) and assigned a weighting coefficient of 0.3-0.7 (the more severe the fault, the lower the weighting coefficient, to avoid high-risk samples misleading the later prediction model, which is the improved DCCNN model). Samples not identified as high-risk samples retain a weight of 1.0 and are not subject to weighting.

[0128] The following explains the weighting coefficients:

[0129] The formula for calculating the weighting factor is:

[0130]

[0131] in, For weighting coefficients, This is a severity index of the fault. This is the distance influence factor.

[0132] Fault Severity Index The calculation formula is:

[0133]

[0134] Where act is the actual value and thr is the fault stage threshold. G For gas concentration, T For hot spot temperature, Q denoted as tanδ, where tanδ is the partial discharge quantity and tanδ is the dielectric loss factor.

[0135] Distance Influence Factor The calculation formula is:

[0136]

[0137] in, D max The maximum distribution distance of the samples. This represents the distance between the sample and the decision boundary.

[0138] When the fault is most severe ( S =1) and the sample is close to the boundary ( K When =0), W =0.3 (lowest weight; the more severe the fault, the lower the coefficient).

[0139] When the fault is minor ( S =0.1) and the sample is close to the upper limit of the threshold ( K When =1), W =0.3+0.4×0.9×1≈0.66 (close to 0.7, which is consistent with the higher coefficient when the risk is low).

[0140] When the fault is moderate ( S =0.5) and the distance is moderate ( K When =0.5), W =0.3+0.4×0.5×0.5=0.4 (intermediate weight, balancing risk and model stability).

[0141] Abnormal samples from real-time acquired dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data in oil-immersed transformer oil are subjected to sample equalization processing. After obtaining the equalized samples, the stability of the equalized samples is judged (in... Figure 3 (Using boundary stability evaluation) If the stability of the equalized samples does not meet the conditions, the equalized samples are subjected to sample equalization again. The indicators for judging the stability of the equalized samples include the Kappa coefficient (measures the classification consistency of the equalized samples; a Kappa coefficient ≥ 0.9 indicates stability), the distribution density of the equalized samples (a distribution density ≥ 0.7 indicates stability), and the proportion of boundary samples (a boundary sample proportion ≤ 15% indicates stability).

[0142] The following provides a detailed explanation of the GAF image channels, ratio feature channels, and physical feature channels:

[0143] A.GAF image channel.

[0144] One-dimensional DGA data is converted into two-dimensional images (64×64) of GASF (Gramian Angular Summation Field) and GADF (Gramian Angular Difference Field) using Gram angle field (GAF) to preserve the temporal evolution of gas concentration. In power engineering, the changing trend of fault gas concentration is more predictive than a single value (e.g., a continuous increase in C2H2 indicates a worsening of the discharge fault), and the two-dimensional image can intuitively present this engineering characteristic. Since the Gram angle field relies on the calculation of inverse trigonometric functions, the first step is to convert the original time series... Normalization to a suitable interval is necessary to ensure numerical stability and avoid polarity reversal. The minimum-maximum normalization method is typically used, scaling the value to the interval [-1, 1] or [0, 1]. The calculation formula is as follows:

[0145]

[0146] in, The values ​​before normalization. The value is the normalized value. , Original time series The maximum value, Original time series The minimum value.

[0147] The normalized time series is transformed to polar coordinates. In this transformation, the normalized values ​​are mapped to polar angles, while the timestamps are mapped to polar radii. The calculation formula is as follows:

[0148]

[0149] in, For the first k The polar angle of each point , , This represents the total length of the sequence.

[0150]

[0151] in, For the first k The polar radius of each point For the first k A timestamp at each point, This represents the total length of the sequence.

[0152] By calculating the sum or difference of trigonometric functions between polar angles and the polar radius, two main Gram angle fields are generated: Gram angle sum field (GASF) and Gram angle difference field (GADF).

[0153] GASF constructs a matrix by calculating the cosine of the sum of every two angles, using the following formula:

[0154]

[0155] in, Let I be the normalized time series vector, and let I be the unit vector.

[0156] GADF constructs a matrix by calculating the sine of the difference between every two angles. The calculation formula is as follows:

[0157]

[0158] B. Ratio characteristic channel.

[0159] Input the 16-dimensional fault sensitivity ratio feature data shown in Table 1, directly connect to the core logic of traditional power prediction methods, and ensure that the subsequent prediction model (improved DCCNN model) is compatible with the existing operation and maintenance system to avoid technical gaps.

[0160] C. Physical characteristic channel.

[0161] Three types of physical parameters are input: hot spot temperature, partial discharge quantity, and dielectric loss factor of insulating oil. The input of the later prediction model (improved DCCNN model) is adapted by Min-Max normalization (mapped to the [0,1] interval). Among them, hot spot temperature reflects the severity of thermal fault, partial discharge quantity reflects the intensity of discharge fault, and dielectric loss factor reflects the aging state of insulation. The three together constitute the physical essential characteristics of oil-immersed transformer faults.

[0162] S3. Based on the sample data, the fault stages of the oil-immersed transformer are divided, and based on the results of the fault stage division of the oil-immersed transformer, the current fault stage of the oil-immersed transformer is determined.

[0163] The fault stage division results for oil-immersed transformers in this step include the fault latency stage, fault development stage, and fault occurrence stage. The fault latency stage, fault development stage, and fault occurrence stage are explained in detail below:

[0164] A. Fault latency stage

[0165] It only produces H2 and CH4, has a hot spot temperature T < 90℃, a partial discharge quantity < 500pC, a dielectric loss factor tanδ < 0.01, and exhibits weak fault characteristics.

[0166] B. Fault Development Stage

[0167] C2H4 and C2H6 surged, hot spot temperature 90℃≤T<150℃, partial discharge quantity 500pC≤Q<1000pC, dielectric loss factor 0.01≤tanδ<0.02, and fault characteristics gradually emerged.

[0168] C. Fault Occurrence Stage

[0169] The presence of C2H2 with a concentration >10 μL / L, hotspot temperature T ≥150℃, partial discharge quantity Q ≥1000 pC, and dielectric loss factor tanδ ≥0.02 indicate that a fault has formed and may worsen.

[0170] S4. Based on the current fault stage of the oil-immersed transformer, the feature vectors of the GAF image channel, the feature vector of the ratio feature channel, and the feature vector of the physical feature channel are fused to obtain the fused channel feature vector.

[0171] Based on the current fault stage of the oil-immersed transformer, the calculation formula for fusing the feature vectors of the GAF image channel, the ratio feature channel, and the physical feature channel is as follows:

[0172]

[0173] in, This is the fused channel feature vector. for function, For GAF image channel feature vectors, The feature vector of the ratio feature channel. For physical feature channels, feature vectors The weights of the feature vectors for the GAF image channels. The weights of the feature vectors for the ratio feature channels are... The weights of the physical feature channel feature vectors. .

[0174] , and It can adaptively adjust according to the current fault stage of the oil-immersed transformer (in Figure 3 Chinese (Weight allocation representation), in this embodiment, the physical feature channel feature vector weights during the fault latency stage. =0.5 (at this point, gas characteristics are not obvious, and physical characteristics are more sensitive), the ratio characteristic channel feature vector weight of the fault development stage. =0.4, weight of GAF image channel feature vector during the fault occurrence stage =0.5 (at this point, the gas timing changes drastically).

[0175] S5. Adaptive feature extraction is performed on the dissolved gas analysis data and physical characteristic data of the oil-immersed transformer oil acquired in real time to determine the parameters of the improved DCCNN model.

[0176] In this step, adaptive feature extraction is performed on the dissolved gas analysis data and physical characteristic data of the oil-immersed transformer oil acquired in real time to determine the parameters of the improved DCCNN model. Specifically, the fault identification model is used to determine the parameters of the improved DCCNN model based on the dissolved gas analysis data and physical characteristic data of the oil-immersed transformer oil acquired in real time.

[0177] The following details the adaptive feature extraction performed on the real-time acquired dissolved gas analysis data and physical characteristic data of oil-immersed transformer oil to determine the parameters of the improved DCCNN model:

[0178] This embodiment specifically uses a fault identification model (e.g., based on logistic regression algorithm) to adaptively extract features from the real-time acquired dissolved gas analysis data and physical characteristic data in oil-immersed transformer oil, determining the parameters of the improved DCCNN model (specifically, determining the void ratio of the improved DCCNN model). The following section discusses the void ratios corresponding to the fault latency stage, fault development stage, and fault occurrence stage. Explanation:

[0179] Fault latency stage: void ratio =1 (standard convolution), with a kernel size of 3×3, to enhance the extraction of weak features;

[0180] Fault development stages: Void rate =3, expand the receptive field to capture multi-scale features;

[0181] Fault occurrence stage: Void rate =5, focusing on capturing global correlations of strong fault characteristics.

[0182] The improved DCCNN model in this step includes:

[0183] Dilated convolutional modules are added between the elements of the standard convolutional kernel in the DCCNN model to expand the receptive field.

[0184] The formula for calculating the dilated convolution module is:

[0185]

[0186] in, For the index of the convolutional layer, For the first convolutional layer feature maps at location The activation value on For the first The kernel weights of the convolutional layer For input data, For the first The bias term of the convolutional layer, Let σ be the void ratio, and σ be the nonlinear activation function. For the first Row index of the convolution kernel of the convolutional layer. For the first Column index of the convolution kernel of the convolutional layer. and These are the x and y coordinates on the feature map.

[0187] S6. Input the fused channel feature vectors into the improved DCCNN model with determined parameters to obtain the fault prediction results of the oil-immersed transformer.

[0188] The fused channel feature vectors are input into an improved DCCNN (IDCCNN, Improved DCCNN) model with predetermined parameters. Figure 3 The IDCCNN model is used to predict the faults of oil-immersed transformers.

[0189] The construction of the improved DCCNN model is explained in detail below:

[0190] A. Convolutional Neural Networks

[0191] Convolutional Neural Networks (CNNs) are a typical deep learning algorithm, usually consisting of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer stacked sequentially. The convolutional layer is the core of a CNN, and its function is feature extraction. The formula for calculating the convolution process is:

[0192]

[0193] in, This is the index of the convolutional layer, representing the current computation level. Layered convolutional networks, For the first convolutional layer feature maps at location The activation value on For the first The kernel weights of the convolutional layer For input data, For the first The bias term of the convolutional layer, where σ is a non-linear activation function. For the first Row index of the convolution kernel of the convolutional layer. For the first Column index of the convolution kernel of the convolutional layer. and Let x and y be the x and y coordinates on the feature map, representing the first x and y coordinates. On the feature map of the convolutional layer The pixel at the location.

[0194] B. Depthwise Separable Convolution (DSC)

[0195] Depthwise separable convolution (DSC) is a technique designed to reduce the computational complexity of CNNs. DSC primarily consists of depthwise convolutions and 1×1 pointwise convolutions. The computational costs of DSC and traditional convolutions are as follows:

[0196]

[0197]

[0198]

[0199] in, This represents the computational cost of convolution operations when using DSC. To calculate the computational cost of convolution operations when using traditional convolution, The size of the convolution kernel. and To output the height and width of the feature map, Input the number of channels. This represents the number of output channels.

[0200] according to It is known that DSC can significantly reduce the computational cost of convolution operations. Therefore, the DSC module is used to reduce the computational complexity of the DCCNN model. Furthermore, to address the death problem of the commonly used non-linear activation function ReLU, GeLU is used instead of ReLU. The GeLU function can maintain stable gradient updates in deep networks. The GeLU function expression is:

[0201]

[0202] in, For gelu functions, Let be the error function. For input variables.

[0203] The principle of C.IDCCNN model

[0204] The DCCNN model combines 2D-CNN (two-dimensional-convolutional neural network) and 1D-CNN (one-dimensional-convolutional neural network). The core design of the DCCNN model is that two independent convolutional sub-networks extract features from different dimensions or modalities, and then achieve feature complementarity through fusion mechanisms such as concatenation and attention weighting. With its stronger feature capture ability and robustness, the DCCNN model performs well in scenarios such as fault prediction.

[0205] To further enhance the DCCNN model's ability to extract long-range dependencies and multi-scale features from oil-immersed transformer fault sequence data, this invention adds a dilated convolution module to the original DCCNN model structure, constructing an improved dual-channel convolutional neural network (IDCCNN). The added dilated convolution module is described in detail below:

[0206] Dilated convolutional modules, also known as spaced convolutional modules, expand the receptive field by adding dilated (spaced) convolutional modules between the elements of the standard convolutional kernel in a DCCNN model. This process does not increase the number of parameters or computational complexity. The formula for calculating a dilated convolutional module is:

[0207]

[0208] in, This is the index of the convolutional layer, representing the current computation level. Layered convolutional networks, For the first convolutional layer feature maps at location The activation value on For the first The kernel weights of the convolutional layer For input data, For the first The bias term of the convolutional layer, The void ratio, It is a positive integer, and σ is a nonlinear activation function. For the first Row index of the convolution kernel of the convolutional layer. For the first Column index of the convolution kernel of the convolutional layer. and Let x and y be the x and y coordinates on the feature map, representing the first x and y coordinates. On the feature map of the convolutional layer The pixel at the location.

[0209] When void ratio When the dilated convolution equals 1, it degenerates into a standard convolution. With... As the value of increases, the receptive field of the convolutional kernel expands exponentially. This mechanism enables the IDCCNN model to capture a wider range of contextual information in deeper layers without losing detail due to pooling operations, making it particularly suitable for analyzing DGA data with complex temporal dependencies.

[0210] The overall structure of the improved DCCNN model includes:

[0211] Input layer: Three-channel feature concatenation;

[0212] Feature extraction layer: 3 layers of depthwise separable convolution (DSC) and 2 layers of dilated convolution. The DSC module splits the traditional convolution into depthwise convolution and 1×1 pointwise convolution, reducing the computational load by 75% (in line with the computing power limitations of power field edge computing equipment). The dilated convolution adjusts the dilation rate in stages.

[0213] Fully connected layers: 2 fully connected layers (256 and 128 hidden neurons), using the GELU activation function (to avoid gradient vanishing caused by the complex features of electrical data);

[0214] Output layer: 5 types of faults (medium and low temperature overheating, high temperature overheating, low energy discharge, high energy discharge and partial discharge) + credibility + gas contribution.

[0215] This embodiment employs the Optuna hyperparameter optimization framework to optimize the hyperparameters of the improved DCCNN model after inputting the data (in... Figure 3 The Optuna hyperparameter optimization framework (referred to as Optuna optimization) integrates advanced sampling algorithms (such as TPE and Tree-structured Parzen Estimator) to dynamically analyze historical "experiment" results during optimization and obtain more promising hyperparameter combinations, thus significantly improving sampling efficiency. Furthermore, the Optuna hyperparameter optimization framework possesses powerful pruning capabilities, employing mechanisms such as asynchronous continuous halving algorithms to monitor intermediate results of ongoing "experiments" in real time and automatically terminate poorly performing experiments, ultimately finding the parameters that optimize the performance of the improved DCCNN model.

[0216] To address the real-time and stability requirements of the power model, the hyperparameter search range set in this embodiment is as follows:

[0217] Learning rate: 0.001-0.01 (adapted to the convergence characteristics of power data);

[0218] Number of iterations: 30-80 rounds (introducing an early stopping mechanism, stopping if the accuracy of the validation set does not improve after 3 rounds);

[0219] Number of convolution kernels: 64-128 (balancing computational power and feature extraction capability);

[0220] The Optuna hyperparameter optimization framework uses the TPE sampling algorithm and pruning mechanism to complete the optimization within 1250s (which can shorten the time by 60% compared to the grid search method) and obtain the optimal hyperparameter combination.

[0221] The fault prediction results for oil-immersed transformers include three types of information: fault type, prediction confidence level, and gas contribution. The following provides a detailed explanation of fault type, prediction confidence level, and gas contribution:

[0222] Fault Types: Clearly output 5 common power faults (medium and low temperature overheating, high temperature overheating, low energy discharge, high energy discharge, and partial discharge), consistent with the power equipment fault classification standard;

[0223] Prediction confidence: The output model (improved DCCNN model) predicts the probability (range [0,1]). A confidence level ≥ 0.9 indicates high reliability, 0.8 ≤ confidence level < 0.9 indicates medium reliability (supplementary monitoring is recommended), and confidence level < 0.8 indicates low reliability (triggers secondary prediction and re-collection of data).

[0224] Gas contribution: Output the influence weight of each characteristic gas on the fault (e.g., C2H2 contribution ≥ 0.7 during high-energy discharge) to clarify the core cause of the fault.

[0225] Based on the fault prediction results of oil-immersed transformers and the maintenance recommendations provided in the "Guidelines for the Maintenance of Power Transformers" DL / T 573-2021, these recommendations directly support on-site practical operations by maintenance personnel. The maintenance recommendations are explained in detail below:

[0226] Medium and low temperature overheating (confidence ≥0.9): It is recommended to monitor the hot spot temperature change within 1 month, check the radiator operation status, and conduct oil quality testing;

[0227] High-energy discharge (reliability ≥ 0.9): Immediately arrange for shutdown and maintenance, focusing on checking the winding insulation condition and tap changer contact, and conduct partial discharge location tests;

[0228] Partial discharge (0.8 ≤ confidence level < 0.9): Increase the oil chromatography sampling frequency to once every 2 hours, monitor the C2H2 concentration change, and initiate maintenance if it continues to rise.

[0229] Example 5:

[0230] Please see Figure 4As shown, the present invention also provides an electronic device 100 for a fault prediction method for oil-immersed transformers; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0231] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the fault prediction method for the oil-immersed transformer described in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0232] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0233] The memory 101 in the electronic device 100 stores multiple instructions to implement a fault prediction method for oil-immersed transformers, and the processor 102 can execute the multiple instructions to achieve the following:

[0234] Real-time acquisition of dissolved gas analysis data in oil of oil-immersed transformers, as well as fault sensitivity ratio characteristic data and physical characteristic data, are recorded as the first data, the second data and the third data, respectively, which together form sample data, and the sample data includes normal samples and abnormal samples; and the fault stages of oil-immersed transformers are divided according to the sample data, and the current fault stage of oil-immersed transformers is determined according to the fault stage division results of oil-immersed transformers.

[0235] The first data retains the time evolution law, the second data extracts the sensitivity features reflecting the faults of oil-immersed transformers, and the third data extracts the physical essence features reflecting the faults of oil-immersed transformers, thus obtaining the first feature vector, the second feature vector, and the third feature vector.

[0236] Based on the current fault stage of the oil-immersed transformer, the first feature vector, the second feature vector, and the third feature vector are fused to obtain the fused channel feature vector.

[0237] The fused channel feature vectors are input into an improved DCCNN model with predetermined parameters to obtain fault prediction results for oil-immersed transformers. The improved DCCNN model is obtained by adding dilated convolution modules between the elements of the standard convolution kernel of the DCCNN model. The dilated convolution modules are used to expand the receptive field, and the parameters are determined based on the first and third data.

[0238] Example 6:

[0239] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0240] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0241] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0242] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0243] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0244] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A fault prediction method for an oil-immersed transformer, characterized in that, Includes the following steps: Real-time acquisition of dissolved gas analysis data in oil-immersed transformer oil, as well as fault sensitivity ratio characteristic data and physical characteristic data, are recorded as the first data, the second data, and the third data, respectively, which together form sample data. The sample data includes normal samples and abnormal samples. The fault stages of the oil-immersed transformer are divided according to the sample data, and the current fault stage of the oil-immersed transformer is determined according to the fault stage division results. The physical characteristic data includes hot spot temperature, partial discharge quantity, and insulating oil dielectric loss factor. The fault stages include fault latency stage, fault development stage, and fault occurrence stage. The first data retains the time evolution law, the second data extracts the sensitivity features reflecting the faults of oil-immersed transformers, and the third data extracts the physical essence features reflecting the faults of oil-immersed transformers, thus obtaining the first feature vector, the second feature vector, and the third feature vector. Based on the current fault stage of the oil-immersed transformer, the first feature vector, the second feature vector, and the third feature vector are fused to obtain the fused channel feature vector. The formula for fusing the first, second, and third eigenvectors is as follows: in, This is the fused channel feature vector. for function, The first eigenvector, This is the second feature vector. The third eigenvector, The weights of the first feature vector. The weights of the second feature vector. The weights of the third feature vector. , and Adaptive adjustments are made based on the current fault stage of the oil-immersed transformer; The fused channel feature vectors are input into an improved DCCNN model with determined parameters to obtain the fault prediction results of the oil-immersed transformer. The improved DCCNN model is obtained by adding dilated convolution modules between the elements of the standard convolution kernel of the DCCNN model. The dilated convolution modules are used to expand the receptive field. The parameters are determined based on the first data and the third data, and the parameter is the porosity.

2. The fault prediction method for oil-immersed transformers according to claim 1, characterized in that, The time evolution law of the first data is retained, the sensitivity features reflecting the fault of the oil-immersed transformer are extracted from the second data, and the physical essence features reflecting the fault of the oil-immersed transformer are extracted from the third data. Before obtaining the first feature vector, the second feature vector and the third feature vector, the abnormal samples in the first data, the second data and the third data are subjected to sample equalization processing to obtain the equalized samples.

3. The fault prediction method for oil-immersed transformers according to claim 2, characterized in that, The outlier samples in the first, second, and third data sets are subjected to sample equalization. After obtaining the equalized samples, the stability of the equalized samples is judged. If the stability of the equalized samples does not meet the conditions, the equalization process is repeated.

4. The fault prediction method for oil-immersed transformers according to claim 2, characterized in that, The abnormal samples include faulty samples and high-risk samples; The process of equalizing outlier samples in the first, second, and third data sets specifically involves: The fault samples in the first, second, and third data are subjected to SMOTE oversampling for sample equalization to obtain synthetic samples. The noise in the synthesized samples was processed using the IDBCS method for sample equalization. High-risk samples in the first, second, and third datasets are weighted to achieve sample equalization.

5. The fault prediction method for oil-immersed transformers according to claim 1, characterized in that, The parameters are determined based on the first data and the third data. Specifically, the parameters of the improved DCCNN model are determined by using a fault identification model to perform adaptive feature extraction on the first data and the third data.

6. The fault prediction method for oil-immersed transformers according to claim 1, characterized in that, The calculation formula for the dilated convolution module is as follows: in, For the index of the convolutional layer, For the first convolutional layer feature maps at location The activation value on For the first The kernel weights of the convolutional layer For input data, For the first The bias term of the convolutional layer, Let σ be the void ratio, and σ be the nonlinear activation function. For the first Row index of the convolution kernel of the convolutional layer. For the first Column index of the convolution kernel of the convolutional layer. and These are the x and y coordinates on the feature map.

7. A fault prediction system for an oil-immersed transformer, characterized in that, include: The data acquisition module is used to acquire dissolved gas analysis data, fault sensitivity ratio characteristic data, and physical characteristic data in real time for oil-immersed transformers. These are denoted as first data, second data, and third data, respectively, and together they form sample data, which includes normal and abnormal samples. The module then divides the fault stages of the oil-immersed transformer based on the sample data, and determines the current fault stage of the transformer based on the fault stage division results. The physical characteristic data includes hot spot temperature, partial discharge quantity, and insulating oil dielectric loss factor. The fault stages include fault latency stage, fault development stage, and fault occurrence stage. The feature vector determination module is used to retain the time evolution law of the first data, extract the sensitivity features reflecting the fault of the oil-immersed transformer from the second data, and extract the physical essence features reflecting the fault of the oil-immersed transformer from the third data, thereby obtaining the first feature vector, the second feature vector and the third feature vector. The fusion module is used to fuse the first feature vector, the second feature vector, and the third feature vector according to the current fault stage of the oil-immersed transformer to obtain the fused channel feature vector. The formula for fusing the first, second, and third eigenvectors is as follows: in, This is the fused channel feature vector. for function, The first eigenvector, This is the second feature vector. The third eigenvector, The weights of the first feature vector. The weights of the second feature vector. The weights of the third feature vector. , and Adaptive adjustments are made based on the current fault stage of the oil-immersed transformer; The prediction module is used to input the fused channel feature vector into an improved DCCNN model with determined parameters to obtain the fault prediction result of the oil-immersed transformer. The improved DCCNN model is obtained by adding a dilated convolution module between the elements of the standard convolution kernel of the DCCNN model. The dilated convolution module is used to expand the receptive field. The parameters are determined based on the first data and the third data, and the parameter is the dilation rate.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the fault prediction method for the oil-immersed transformer according to any one of claims 1 to 6.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the fault prediction method for the oil-immersed transformer according to any one of claims 1 to 6.

Images